antisrdy / el_nino

Predict temperature in the Pacific El Nino zone on a 6-month ahead basis

3-day El Nino data challenge

Goal was to predict the temperature in the Pacific El Nino zone on a 6-month ahead basis.

File description

• The notebook named "el_nino..." sets up the problem and gives a piece of context
• The notebook named "Antoine..." presents different steps taken from time 0, what did not work, what looked promising but did not work, and what worked the best
• ts_feature_extractor.py and regressor.py enable to score top 3 on the public leaderboard
• Models can be easily trained, just running user_test_submission.py

Main lessons

What worked best to predict temperatures at time t+6:

• Feature engineering:
  • Considering worldwide temperatures at time t
  • Considering average temperature in the El Nino zone at time t - 6 (standing for the same month to predict, but one year before)
• Prediction step was an average of:
  • XGBoost in order to extract most relevant geographic zones, which can be assimilated to feature selection. Best parameters were estimated using CV (see the notebook for more details)
  • Linear regression with Lasso penalization

Below is a Figure that shows how much a geographic zone leverages the temperature in the El-Niño zone:

Requirements

• XGBoost: http://xgboost.readthedocs.io/en/latest/build.html
• Some utilities. If you're using Anaconda: conda install xarray dask netCDF4 bottleneck basemap

Usage

Open a terminal

git clone https://github.com/antisrdy/el_nino
cd el_nino
python user_test_submission.py